NightSplat: Learning 3D Scene Reconstruction from Nighttime Driving Videos

Overview

Nighttime driving scenes remain extremely challenging for neural rendering and 3D reconstruction methods due to low-light conditions, sensor noise, motion blur, lens flare artifacts, and highly dynamic lighting from vehicles and street lamps.

We present NightSplat, a Gaussian Splatting framework specifically designed for high-fidelity nighttime scene reconstruction and novel view synthesis from automotive driving sequences.

Overview. We represent the scene as a set of 3D Gaussians rendered in raw sensor space, modeling long-exposure blur, moving light sources, lens flares, and acquisition noise to faithfully reconstruct nighttime driving scenes.

Qualitative Results

NightSplat reconstructs sharper nighttime scene details and significantly improves the realism of nighttime driving scenes, particularly around dynamic lighting and flare-heavy regions.

Qualitative comparison on NuScenes. NightSplat (ours) reconstructs sharper scene details with more realistic lighting compared to prior methods.

Method

NightSplat extends Gaussian Splatting to nighttime driving conditions through four key contributions:

Exposure-aware temporal rendering
Raw sensor-space reconstruction with noise modeling
Lens flare disentanglement
Dynamic moving-light modeling

Light modeling. Each dynamic light source is represented by an appearance embedding and relative position. Light features are processed by a neural network alongside splatted scene features and surface normals to predict illumination effects. A separate lightweight CNN estimates lens flare contributions from saturated scene areas.

BibTeX

@article{ramazzina2026nightsplat,
  title={NightSplat: Learning 3D Scene Reconstruction from Nighttime Driving Videos},
  author={Ramazzina, Andrea and Gasperini, Stefano and Bijelic, Mario and Heide, Felix and Tombari, Federico},
  journal={IEEE Robotics and Automation Letters (RA-L)},
  year={2026}
}